CN114860462B - Intelligent computing resource distribution system and method for double-path rack-mounted server - Google Patents

Intelligent computing resource distribution system and method for double-path rack-mounted server Download PDF

Info

Publication number
CN114860462B
CN114860462B CN202210810557.6A CN202210810557A CN114860462B CN 114860462 B CN114860462 B CN 114860462B CN 202210810557 A CN202210810557 A CN 202210810557A CN 114860462 B CN114860462 B CN 114860462B
Authority
CN
China
Prior art keywords
computing resource
vector
feature vector
corrected
feature
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202210810557.6A
Other languages
Chinese (zh)
Other versions
CN114860462A (en
Inventor
武旻
刘琦
张一鸣
施铮
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing Pudao Electronic Technology Co ltd
Original Assignee
Nanjing Pudao Electronic Technology Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing Pudao Electronic Technology Co ltd filed Critical Nanjing Pudao Electronic Technology Co ltd
Priority to CN202210810557.6A priority Critical patent/CN114860462B/en
Publication of CN114860462A publication Critical patent/CN114860462A/en
Application granted granted Critical
Publication of CN114860462B publication Critical patent/CN114860462B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46—Multiprogramming arrangements
    • G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
    • G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
    • G06F9/5027—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G06F18/24155—Bayesian classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention relates to the field of intelligent allocation of server resources, and particularly discloses an intelligent allocation system and an intelligent allocation method for computing resources of a two-way rack server.

Description

Intelligent computing resource distribution system and method for two-way rack-mounted server
Technical Field
The present invention relates to the field of intelligent allocation of server resources, and more particularly, to an intelligent allocation system and an intelligent allocation method for computing resources of a two-way rack server.
Background
The server is one of computers, which runs faster, has higher load and is more expensive than a common computer, and provides calculation or application services for other clients (such as terminals of a PC, a smart phone, an ATM and the like, and even large equipment such as a train system and the like) in a network. The popular server architecture forms in the market at present are tower type, rack type and blade type. Rack servers, as well as modified rack servers and even tower servers, will continue to be the mainstream of the server market.
However, at present, the pooling and scheduling of the computing resources of the server are not perfect enough, so that the data is extruded and lost, the scheduling is not suitable, the stability and the later storage of the data are influenced, and the resource loss is caused indirectly.
In the conventional scheme allocation of computing resources of a server, the computing resources are usually allocated based on the computing requirements of clients, and a mode of allocating resources preferentially by a first-come client is usually adopted, but this allocation mode may cause that other clients cannot obtain enough computing resources when sharing the server, and the server itself also needs to maintain normal operation, such as data transmission, data storage, and data coordination, that is, when the server allocates computing resources, the server needs to consider the remaining computing resource conditions of the server itself in addition to the requirements of all the clients for computing resources as a whole.
Therefore, an optimized intelligent allocation scheme for computing resources of a server is expected.
Disclosure of Invention
The present application is proposed to solve the above-mentioned technical problems. The embodiment of the application provides an intelligent computing resource distribution system of a two-way rack server and a distribution method thereof, implicit associated characteristics of the required computing resource amount of a client, the residual computing resource amount of the server and the computing resource amount distributed by the client in a time dimension are mined through a deep neural network model, the rationality judgment of the computing resource distribution proportion of the server is carried out by integrating characteristic information of the three components, and through the mode, the computing resource distribution scheme is intelligently adjusted according to the overall situation of the required computing resource amount of the client and the residual computing resource condition of the server, so that the distribution rationality and the utilization efficiency of the computing resources of the server are improved.
According to one aspect of the application, a computing resource intelligent distribution system of a two-way rack server is provided, which comprises:
the client computing request acquisition module is used for acquiring computing resource requests of a plurality of clients at a plurality of preset time points including the current time point, wherein the computing resource requests comprise the required computing resource amount;
the computing request data encoding module is used for inputting the required computing resource quantity in the computing resource requests of the plurality of clients at each preset time point into the context encoder to obtain a global demand characteristic vector corresponding to each preset time point;
the global mean processing module is used for calculating the global mean of the global demand characteristic vector corresponding to each preset time point to obtain a global calculation demand characteristic value corresponding to each preset time point;
the calculation request time sequence coding module is used for arranging the global calculation demand characteristic values corresponding to the preset time points into input vectors according to the time dimension and then obtaining calculation resource demand characteristic vectors through a first time sequence coder comprising a one-dimensional convolutional layer;
the residual computing resource amount acquisition module is used for acquiring the residual computing resource amount of the servers at a plurality of preset time points including the current time point;
the residual computing resource amount coding module is used for enabling the residual computing resource amount of the server at a plurality of preset time points including the current time point to pass through a second time sequence encoder comprising a one-dimensional convolutional layer so as to obtain a residual computing resource feature vector;
the distribution data acquisition module is used for acquiring the amount of computing resources distributed to the plurality of clients by the server at the current time point;
the distribution data encoding module is used for enabling the computing resource amount distributed to the plurality of clients by the server at the current time point to pass through a third time sequence encoder containing a one-dimensional convolution layer so as to obtain a current computing resource distribution characteristic vector;
a Bayesian fusion module for fusing the computing resource demand feature vector, the current computing resource allocation feature vector and the residual computing resource feature vector using a Bayesian probability model to obtain a classification feature vector; and
and the distribution scheme generation module is used for enabling the classification characteristic vectors to pass through a classifier to obtain a classification result, and the classification result is used for indicating whether the calculation resource distribution proportion of the server at the current time point needs to be adjusted or not.
According to another aspect of the application, an allocation method of a computing resource intelligent allocation system of a two-way rack server comprises the following steps:
acquiring computing resource requests of a plurality of clients at a plurality of preset time points including a current time point, wherein the computing resource requests comprise required computing resource amount;
inputting the required computing resource amount in the computing resource requests of the plurality of clients at each preset time point into a context encoder to obtain a global demand characteristic vector corresponding to each preset time point;
calculating a global mean value of the global demand characteristic vector corresponding to each preset time point to obtain a global calculation demand characteristic value corresponding to each preset time point;
arranging the global computation demand characteristic values corresponding to the preset time points into input vectors according to time dimension, and then obtaining computation resource demand characteristic vectors through a first time sequence encoder comprising a one-dimensional convolutional layer;
acquiring the residual computing resource amount of the server at a plurality of preset time points including the current time point;
the residual computing resource amount of the server at a plurality of preset time points including the current time point is processed by a second time sequence encoder comprising a one-dimensional convolution layer to obtain a residual computing resource feature vector;
acquiring the amount of computing resources distributed to the plurality of clients by the server at the current time point;
the computing resource amount distributed to the plurality of clients by the server at the current time point is processed by a third time sequence encoder containing a one-dimensional convolutional layer to obtain a current computing resource distribution characteristic vector;
fusing the computing resource demand feature vector, the current computing resource allocation feature vector and the residual computing resource feature vector by using a Bayesian probability model to obtain a classification feature vector; and
and passing the classified feature vector through a classifier to obtain a classification result, wherein the classification result is used for indicating whether the calculation resource distribution proportion of the server at the current time point needs to be adjusted or not.
Compared with the prior art, the computing resource intelligent distribution system of the two-way rack-mounted server and the distribution method thereof, the method carries out implicit correlation characteristic mining on the required computing resource quantity of the client, the residual computing resource quantity of the server and the computing resource quantity distributed by the client in a time dimension through a deep neural network model, the rationality judgment of the allocation proportion of the computing resources of the server is carried out by integrating the feature information of the three, in the fusion process of the characteristics, the excessive fragmentation of the decision area in the classification target domain caused by the characteristic value set due to the characteristic value outside the distribution of the set is avoided by carrying out the structural understanding of the information rule on the characteristic value and the class condition to which the characteristic value belongs, and then, the constraint that each feature is distributed to a probabilistic classification target is realized, so that the classification accuracy is improved.
Drawings
The above and other objects, features and advantages of the present application will become more apparent by describing in more detail embodiments of the present application with reference to the attached drawings. The accompanying drawings are included to provide a further understanding of the embodiments of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings, like reference numbers generally indicate like parts or steps.
FIG. 1 is a block diagram of a computing resource intelligent distribution system of a two-way rack server according to an embodiment of the present application.
Fig. 2 is a block diagram of a bayesian fusion module in an intelligent computing resource allocation system of a two-way rack server according to an embodiment of the present application.
FIG. 3 is a flowchart illustrating an allocation method of a computing resource intelligent allocation system of a two-way rack server according to an embodiment of the present disclosure.
FIG. 4 is a schematic diagram illustrating an architecture of an allocation method of an intelligent computing resource allocation system of a two-way rack server according to an embodiment of the present disclosure.
Detailed Description
Hereinafter, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. It should be apparent that the described embodiments are only a few embodiments of the present application, and not all embodiments of the present application, and it should be understood that the present application is not limited to the example embodiments described herein.
Overview of scenes
As mentioned above, the server is one of computers, which runs faster, has higher load and is more expensive than a common computer, and provides computing or application services for other clients (such as a terminal like a PC, a smart phone, an ATM, and even a large-scale device like a train system) in the network. The popular server architecture forms in the market at present are tower type, rack type and blade type. Rack servers, as well as modified rack servers and even tower servers, will continue to be the mainstream of the server market.
However, at present, the pooling and scheduling of the computing resources of the server are not perfect enough, so that the data is extruded and lost, the scheduling is not suitable, the stability and the later storage of the data are influenced, and the resource loss is caused indirectly. Therefore, an optimized intelligent allocation scheme for computing resources of a server is expected.
Accordingly, in the conventional solution allocation of computing resources of a server, the computing resources are usually allocated based on the computing requirements of clients, and a mode of allocating the resources preferentially by a first-come client is usually adopted, but this allocation mode may cause that other clients cannot obtain enough computing resources when sharing the server, and the server itself also needs to maintain normal operations, such as data transmission, data storage, and data coordination, that is, when the server allocates the computing resources, in addition to considering the requirements of all the clients for the computing resources as a whole, the remaining computing resources of the server itself need to be considered.
Based on the above, the inventor of the present application considers that, starting from the amount of computing resources required by the client, the amount of computing resources remaining in the server, and the amount of computing resources allocated by the client, mining the correlation characteristics by using a deep neural network model, and classifying the correlation characteristics by using a classifier to obtain the rationality of the allocation ratio of the computing resources of the server. In this way, the loss of resources can be avoided while ensuring that the computing resources of the server are reasonably allocated.
Specifically, in the technical solution of the present application, first, computing resource requests of a plurality of clients at a plurality of predetermined time points including a current time point are obtained, where the computing resource requests include a required amount of computing resources. It should be understood that, considering the correlation between the required computing resource amounts in the computing resource requests of the clients at the predetermined time points, the context encoder, for example, based on the Bert model of the converter, is used to encode the required computing resource amounts in the computing resource requests of the clients at the predetermined time points, so as to extract the global-based high-dimensional semantic features between the required computing resource amounts to better characterize the computing resource requirement features of the clients, thereby obtaining a plurality of client feature vectors. In this way, the plurality of client-side feature vectors are cascaded to obtain the global demand feature vector corresponding to each of the predetermined time points in the time dimension.
Further, a global mean of the global demand feature vectors corresponding to the predetermined time points may be calculated to obtain a global calculation demand feature value having global sample dimension associated feature information corresponding to the predetermined time points. Then, considering that the global computation demand characteristic value of each predetermined time point has dynamic change characteristics in the time dimension, in order to extract such change characteristics in the time dimension, the global computation demand characteristic values corresponding to each predetermined time point are further arranged into an input vector according to the time dimension, so as to integrate the characteristic information of each sample in the time dimension, and then pass through a first time sequence encoder comprising a one-dimensional convolutional layer to obtain a computation resource demand characteristic vector. In one example, the time sequence encoder is composed of full connection layers and one-dimensional convolution layers which are alternately arranged, and the correlation of each global calculation requirement characteristic value on the time sequence dimension is extracted through one-dimensional convolution coding, and high-dimensional implicit characteristics of each global calculation requirement characteristic value are extracted through full connection coding.
It should be understood that, in the technical solution of the present application, in order to judge the rationality of the allocation of the computing resources based on the three of the required computing resource amount of the client, the remaining computing resource amount of the server, and the computing resource amount allocated by the client, it is necessary to perform deep feature mining on the associated feature information of the three, and for the remaining computing resource amount of the server and the computing resource amount allocated by the client, first, the remaining computing resource amount of the server at a plurality of predetermined time points including the current time point and the computing resource amount allocated by the server to the plurality of clients at the current time point are obtained. Then, considering that the remaining computing resource amount of the server in the time sequence dimension or the computing resource amount allocated by the server to the plurality of clients at the current time point has associated implicit feature information with dynamics in time, a time sequence encoder is adopted to extract implicit dynamic associated features of the server and the clients in time sequence respectively, so as to obtain a remaining computing resource feature vector and a current computing resource allocation feature vector.
Further, considering that the current computing resource allocation feature vector is used as the prior probability, the technical solution of the present application aims to update the prior probability to obtain the posterior probability when there is new evidence, i.e. when there are new computing resource requests of the plurality of clients. Then, according to a bayesian formula, a posterior probability is a prior probability multiplied by an event probability divided by an evidence probability, and therefore, in the technical scheme of the present application, a bayesian probability model is used to fuse the computing resource demand feature vector, the current computing resource allocation feature vector and the residual computing resource feature vector to obtain a classification feature vector, wherein the current computing resource allocation feature vector is used as a prior, the residual computing resource feature vector is used as an event, and the computing resource demand feature vector is used as an evidence. In this way, the classification feature vector can be passed through a classifier to obtain a classification result indicating whether the allocation proportion of the computing resources of the server at the current time point needs to be adjusted.
However, considering that the computing resource demand feature vector, the current computing resource allocation feature vector, and the remaining computing resource feature vector need to be mapped to the probability space before using the bayesian probability model, when linear mapping such as maximum normalization is performed, constraint on the feature distribution expressed by the feature vector to the probabilistic classification target cannot be realized, and thus the classification effect of the posterior probability vector calculated by using the bayesian probability model is affected.
Therefore, before using the bayesian probabilistic model, class-conditional boundary constraints are first performed on the computing resource demand feature vector, the current computing resource allocation feature vector, and the remaining computing resource feature vector, specifically:
Figure DEST_PATH_IMAGE001
f i1 、f i2 andf i3 respectively, a feature value for each position of the computing resource demand feature vector, the current computing resource allocation feature vector, and the remaining computing resource feature vector.
Therefore, through the structural understanding of the information rules of the characteristic values and the class conditions to which the characteristic values belong, the excessive fragmentation of the characteristic value set in a decision area in a classification target domain caused by the characteristic values outside the distribution of the set is avoided, and the characteristic distribution expressed by the characteristic value set of the characteristic vector has a steady conditional class boundary, so that the constraint of each characteristic distribution to a probabilistic classification target is realized, the classification effect of the classification characteristic vector obtained by a Bayesian probability model is improved, and the accuracy of the rationality judgment of the calculation resource distribution proportion of the server at the current time point is improved.
Based on this, the present application provides a computing resource intelligent distribution system of a two-way rack server, which includes: the client computing request acquisition module is used for acquiring computing resource requests of a plurality of clients at a plurality of preset time points including the current time point, wherein the computing resource requests comprise the required computing resource amount; the computing request data encoding module is used for inputting the required computing resource quantity in the computing resource requests of the clients at each preset time point into a context encoder so as to obtain a global demand characteristic vector corresponding to each preset time point; the global mean processing module is used for calculating the global mean of the global demand characteristic vector corresponding to each preset time point to obtain a global calculation demand characteristic value corresponding to each preset time point; the calculation request time sequence coding module is used for arranging the global calculation demand characteristic values corresponding to the preset time points into input vectors according to the time dimension and then obtaining calculation resource demand characteristic vectors through a first time sequence coder comprising a one-dimensional convolutional layer; the residual computing resource amount acquisition module is used for acquiring the residual computing resource amount of the servers at a plurality of preset time points including the current time point; the residual computing resource amount coding module is used for enabling the residual computing resource amount of the server at a plurality of preset time points including the current time point to pass through a second time sequence coder including a one-dimensional convolution layer so as to obtain a residual computing resource feature vector; the distribution data acquisition module is used for acquiring the amount of computing resources distributed to the plurality of clients by the server at the current time point; the distribution data encoding module is used for enabling the computing resource amount distributed to the plurality of clients by the server at the current time point to pass through a third time sequence encoder containing a one-dimensional convolution layer so as to obtain a current computing resource distribution characteristic vector; a Bayesian fusion module for fusing the computing resource demand feature vector, the current computing resource allocation feature vector and the residual computing resource feature vector using a Bayesian probability model to obtain a classification feature vector; and the distribution scheme generation module is used for enabling the classification characteristic vectors to pass through a classifier to obtain a classification result, and the classification result is used for indicating whether the calculation resource distribution proportion of the server at the current time point needs to be adjusted or not.
Having described the general principles of the present application, various non-limiting embodiments of the present application will now be described with reference to the accompanying drawings.
Exemplary System
FIG. 1 illustrates a block diagram of a two-way rack server computing resource intelligent allocation system according to an embodiment of the present application. As shown in fig. 1, the intelligent computing resource allocation system 200 for a two-way rack server according to an embodiment of the present application includes: a client computing request collecting module 210, configured to obtain computing resource requests of multiple clients at multiple predetermined time points including a current time point, where the computing resource requests include a required computing resource amount; a computation request data encoding module 220, configured to input a required amount of computation resources in computation resource requests of the multiple clients at each of the predetermined time points into a context encoder to obtain a global demand feature vector corresponding to each of the predetermined time points; a global mean processing module 230, configured to calculate a global mean of the global demand feature vectors corresponding to the respective predetermined time points to obtain a global calculation demand feature value corresponding to the respective predetermined time points; a calculation request time sequence coding module 240, configured to arrange the global calculation requirement characteristic values corresponding to the predetermined time points into an input vector according to a time dimension, and then obtain a calculation resource requirement characteristic vector through a first time sequence encoder including a one-dimensional convolutional layer; a remaining computing resource amount acquiring module 250, configured to acquire remaining computing resource amounts of the servers at multiple predetermined time points including the current time point; a residual computing resource amount encoding module 260, configured to pass the residual computing resource amounts of the servers at multiple predetermined time points including the current time point through a second time sequence encoder including a one-dimensional convolutional layer to obtain a residual computing resource feature vector; a distribution data acquisition module 270, configured to obtain the amount of computing resources distributed by the server to the multiple clients at the current time point; the allocation data encoding module 280 is configured to pass the amount of the computing resources allocated by the server to the plurality of clients at the current time point through a third sequential encoder including a one-dimensional convolutional layer to obtain a current computing resource allocation feature vector; a bayesian fusion module 290 for fusing the computing resource requirement feature vector, the current computing resource allocation feature vector and the remaining computing resource feature vector using a bayesian probability model to obtain a classification feature vector; and an allocation scheme generating module 300, configured to pass the classification feature vector through a classifier to obtain a classification result, where the classification result is used to indicate whether the allocation proportion of the computing resources of the server at the current time point needs to be adjusted.
Specifically, in this embodiment of the present application, the client computation request collecting module 210 and the computation request data encoding module 220 are configured to obtain computation resource requests of multiple clients at multiple predetermined time points including a current time point, where the computation resource requests include a required computation resource amount, and input the required computation resource amount in the computation resource requests of the multiple clients at each of the predetermined time points into a context encoder to obtain a global demand feature vector corresponding to each of the predetermined time points. As mentioned above, in the conventional solution allocation of computing resources of a server, computing resources are usually allocated based on computing requirements of clients, and a mode of allocating resources preferentially by first-come users is usually adopted, but this allocation mode may cause that other clients cannot obtain enough computing resources when sharing the server, and the server itself needs to maintain normal operation, such as data transmission, data storage, and data coordination, that is, when the server allocates computing resources, it needs to consider the remaining computing resources of the server itself in addition to the requirements of all the clients for computing resources as a whole. Therefore, in the technical scheme of the application, the mining of the association characteristics is performed by using the deep neural network model from the amount of the required computing resources of the client, the amount of the remaining computing resources of the server and the amount of the computing resources allocated by the client, and then the classification is performed by using the classifier so as to obtain the rationality of the computing resource allocation proportion of the server. Thus, the loss of resources can be avoided under the condition of ensuring that the computing resources of the server are reasonably distributed.
That is, specifically, in the technical solution of the present application, first, computing resource requests of a plurality of clients at a plurality of predetermined time points including a current time point are obtained, where the computing resource requests include a required amount of computing resources. It should be understood that, in consideration of the correlation between the required computing resource amounts in the computing resource requests of the clients at the predetermined time points, in the technical solution of the present application, a context encoder, for example, based on a Bert model of a converter, is used to encode the required computing resource amounts in the computing resource requests of the clients at the predetermined time points, so as to extract a global-based high-dimensional semantic feature between the required computing resource amounts to better characterize the computing resource requirement features of the clients, thereby obtaining a plurality of client feature vectors. In this way, the plurality of client-side feature vectors are cascaded to obtain the global demand feature vector corresponding to each of the predetermined time points in the time dimension.
More specifically, in this embodiment of the present application, the computation request data encoding module includes: a context encoding unit, configured to perform global context-based semantic encoding on the required amount of computing resources in the computing resource requests of the multiple clients at each of the predetermined time points using a Bert model based on a converter of the context encoder to obtain multiple client feature vectors; and the cascading unit is used for cascading the plurality of client characteristic vectors to obtain the global demand characteristic vector corresponding to each preset time point.
Specifically, in this embodiment of the present application, the global mean processing module 230 and the calculation request time sequence encoding module 240 are configured to calculate a global mean of the global demand feature vectors corresponding to each of the predetermined time points to obtain a global calculation demand feature value corresponding to each of the predetermined time points, and arrange the global calculation demand feature values corresponding to each of the predetermined time points into an input vector according to a time dimension, and then obtain a calculation resource demand feature vector through a first time sequence encoder including a one-dimensional convolutional layer. That is, in the technical solution of the present application, further, a global mean of the global demand feature vector corresponding to each of the predetermined time points may be calculated to obtain a global calculation demand feature value having global sample dimension associated feature information corresponding to each of the predetermined time points. Then, considering that the global computation demand characteristic value of each predetermined time point has dynamic change characteristics in the time dimension, in order to extract such change characteristics in the time dimension, the global computation demand characteristic values corresponding to each predetermined time point are further arranged into an input vector according to the time dimension, so as to integrate the characteristic information of each sample in the time dimension, and then pass through a first time sequence encoder comprising a one-dimensional convolutional layer to obtain a computation resource demand characteristic vector. Accordingly, in a specific example, the time sequence encoder is composed of a full connection layer and a one-dimensional convolution layer which are alternately arranged, and the correlation of each global computation requirement characteristic value in the time sequence dimension is extracted through one-dimensional convolution coding, and the high-dimensional implicit characteristic of each global computation requirement characteristic value is extracted through full connection coding.
More specifically, in this embodiment of the present application, the computation request timing coding module includes: the first arrangement unit is used for arranging the global calculation demand characteristic values corresponding to the preset time points into input vectors according to time dimensions; a first full-concatenation coding unit, configured to perform full-concatenation coding on the input vector using a full-concatenation layer of the first timing coder according to a following formula to extract a high-dimensional implicit feature of a feature value of each position in the input vector, where the formula is:
Figure 755129DEST_PATH_IMAGE002
where X is the input vector, Y is the output vector, W is the weight matrix, B is the offset vector,
Figure DEST_PATH_IMAGE003
represents a matrix multiplication; a first one-dimensional convolution coding unit, configured to perform one-dimensional convolution coding on the input vector by using a one-dimensional convolution layer of the first time sequence encoder according to the following formula to extract a high-dimensional implicit correlation feature between feature values of each position in the input vector, where the formula is:
Figure 347784DEST_PATH_IMAGE004
wherein the content of the first and second substances,ais a convolution kernelxA width in the direction,FIs a convolution kernel parameter vector,GIs a matrix of local vectors operating with a convolution kernel,wthe size of the convolution kernel.
Specifically, in this embodiment of the present application, the residual computing resource amount acquiring module 250 and the residual computing resource amount encoding module 260 are configured to acquire the residual computing resource amounts of the servers at the multiple predetermined time points including the current time point, and pass the residual computing resource amounts of the servers at the multiple predetermined time points including the current time point through a second time sequence encoder including a one-dimensional convolution layer to obtain a residual computing resource feature vector. It should be understood that, in the technical solution of the present application, in order to determine the rationality of the allocation of the computing resources based on the three of the required computing resource amount of the client, the remaining computing resource amount of the server, and the computing resource amount allocated by the client, it is necessary to perform deep feature mining on the associated feature information of the three, and for the remaining computing resource amount of the server, first, the remaining computing resource amounts of the server at a plurality of predetermined time points including the current time point are obtained. Then, considering the remaining computing resource amount of the server in the time sequence dimension, the server has associated implicit feature information with dynamics in time, so in the technical scheme of the application, a time sequence encoder is adopted to extract the implicit dynamic associated features of the server in the time sequence, and accordingly, the remaining computing resource feature vector is obtained.
More specifically, in this embodiment, the module for encoding the amount of remaining computing resources includes: the second arrangement unit is used for arranging the residual computing resource quantity of the servers at a plurality of preset time points including the current time point into residual input vectors according to the time dimension; a second fully-concatenated coding unit for using the fully-concatenated layers of the second sequential encoder asPerforming full-concatenation coding on the residual input vector to extract high-dimensional implicit features of feature values of all positions in the residual input vector by using the following formula:
Figure 163293DEST_PATH_IMAGE002
where X is the input vector, Y is the output vector, W is the weight matrix, B is the offset vector,
Figure 575820DEST_PATH_IMAGE003
represents a matrix multiplication; a second one-dimensional convolution coding unit, configured to perform one-dimensional convolution coding on the residual input vector by using a one-dimensional convolution layer of the second time-series encoder according to the following formula to extract a high-dimensional implicit correlation feature between feature values of each position in the residual input vector, where the formula is:
Figure DEST_PATH_IMAGE005
wherein the content of the first and second substances,ais a convolution kernelxA width in the direction,FIs a convolution kernel parameter vector,GIs a matrix of local vectors operating with a convolution kernel,wthe size of the convolution kernel.
Specifically, in this embodiment of the present application, the allocation data collecting module 270 and the allocation data encoding module 280 are configured to obtain the amount of the computing resources allocated by the server to the multiple clients at the current time point, and pass the amount of the computing resources allocated by the server to the multiple clients at the current time point through a third timing encoder that includes a one-dimensional convolutional layer to obtain a current computing resource allocation feature vector. It should be understood that, in the technical solution of the present application, in order to determine the rationality of the allocation of the computing resources based on the three of the required computing resource amount of the client, the remaining computing resource amount of the server, and the computing resource amount allocated by the client, it is necessary to perform deep feature mining on the associated feature information of the three, and for the computing resource amount allocated by the client, first, the computing resource amount allocated by the server to the plurality of clients at the current time point is obtained. Then, considering that the amount of computing resources allocated to the plurality of clients by the server at the current time point in the time sequence dimension has associated implicit feature information that is dynamic in time, in the technical scheme of the present application, a time sequence encoder is adopted to extract the implicit dynamic associated features of the server at the time sequence, so as to obtain the current computing resource allocation feature vector.
More specifically, in this embodiment of the present application, the allocation data encoding module includes: a third arrangement unit, configured to arrange, according to a time dimension, the amount of computing resources allocated by the server to the plurality of clients as an allocation input vector; a third full-concatenation coding unit, configured to perform full-concatenation coding on the allocation input vector using a full-concatenation layer of the third time sequence encoder according to a following formula to extract a high-dimensional implicit feature of a feature value at each position in the allocation input vector, where the formula is:
Figure 134978DEST_PATH_IMAGE002
where X is the input vector, Y is the output vector, W is the weight matrix, B is the offset vector,
Figure 316560DEST_PATH_IMAGE003
represents a matrix multiplication; a third one-dimensional convolutional coding unit, configured to perform one-dimensional convolutional coding on the allocation input vector using a one-dimensional convolutional layer of the third time-series encoder according to the following formula to extract high-dimensional implicit correlation features between feature values of each position in the allocation input vector, where the formula is:
Figure 506233DEST_PATH_IMAGE006
wherein the content of the first and second substances,ais a convolution kernelxA width in the direction,FIs a convolution kernel parameter vector,GIs a matrix of local vectors operating with a convolution kernel,wthe size of the convolution kernel.
Specifically, in this embodiment of the present application, the bayesian fusion module 290 is configured to fuse the computing resource requirement feature vector, the current computing resource allocation feature vector, and the remaining computing resource feature vector by using a bayesian probability model to obtain a classification feature vector. It should be appreciated that further, in view of using the current computing resource allocation feature vector as a prior probability, the technical solution in the present application aims to update the prior probability to obtain a posterior probability when there is new evidence, i.e. when there are new computing resource requests of the plurality of clients. Then, according to a bayesian formula, a posterior probability is a prior probability multiplied by an event probability divided by an evidence probability, and therefore, in the technical scheme of the present application, a bayesian probability model is used to fuse the computing resource demand feature vector, the current computing resource allocation feature vector and the residual computing resource feature vector to obtain a classification feature vector, wherein the current computing resource allocation feature vector is used as a prior, the residual computing resource feature vector is used as an event, and the computing resource demand feature vector is used as an evidence. In this way, the classification feature vector can be passed through a classifier to obtain a classification result indicating whether the allocation proportion of the computing resources of the server at the current time point needs to be adjusted.
More specifically, in an embodiment of the present application, the bayesian fusion module includes: firstly, class-condition-based boundary constraint is respectively carried out on the computing resource demand characteristic vector, the current computing resource allocation characteristic vector and the residual computing resource characteristic vector to obtain a corrected computing resource demand characteristic vector, a corrected current computing resource allocation characteristic vector and a corrected residual computing resource characteristic vector. It should be understood that, considering that the computing resource demand feature vector, the current computing resource allocation feature vector and the remaining computing resource feature vector need to be mapped to the probability space before using the bayesian probability model, when linear mapping such as maximum value normalization is performed, constraint on feature distribution expressed by the feature vector to a probabilistic classification target cannot be realized, so that the classification effect of the posterior probability vector calculated by using the bayesian probability model is affected. Therefore, before using the bayesian probability model, class condition boundary constraints are firstly performed on the feature vector of the computing resource demand, the feature vector of the current computing resource allocation, and the feature vector of the remaining computing resource, specifically:
Figure DEST_PATH_IMAGE007
whereinf i1 、f i2 Andf i3 respectively, a feature value for each position of the computing resource demand feature vector, the current computing resource allocation feature vector, and the remaining computing resource feature vector.
Then, the corrected current computing resource allocation feature vector is used as a prior probability vector, the corrected residual computing resource feature vector is used as an event probability vector, and the computing resource demand feature vector is used as an evidence probability vector, and the classified feature vector is obtained by multiplying the corrected current computing resource allocation feature vector by the corrected residual computing resource feature vector according to the position point and then dividing the position point by the computing resource demand feature vector. Accordingly, in a specific example, the position-by-position point multiplication of the corrected current computing resource allocation feature vector and the corrected remaining computing resource feature vector is divided by the computing resource demand feature vector to obtain the classification feature vector according to the following formula:
Figure 186482DEST_PATH_IMAGE008
whereinf i2 ' is the value of each position in the corrected current computing resource allocation feature vector,f i3 ' andf i1 ' are the corrected residual computing resource feature vector and the corrected computing resource requirement, respectivelyThe value of each position in the feature vector, andf is ' is the value of each location in the classification feature vector.
It should be understood that by performing the structural understanding of the information rules on the eigenvalues and the class conditions to which the eigenvalues belong, excessive fragmentation of the decision region of the eigenvalue set in the classification target domain due to the out-of-distribution eigenvalues of the set is avoided, so that the eigenvalue set of the eigenvector expresses the eigenvalue distribution with a robust conditioned class boundary, thereby realizing the constraint of each eigenvalue distribution to the probabilistic classification target, improving the classification effect of the classification eigenvector obtained by the bayesian probability model, and further improving the accuracy of the rationality judgment of the calculation resource distribution proportion of the server at the current time point.
Fig. 2 is a block diagram of a bayesian fusion module in an intelligent computing resource allocation system of a two-way rack server according to an embodiment of the present application. As shown in fig. 2, the bayesian fusion module 290 includes: a feature distribution correction unit 291, configured to perform class-based boundary constraint on the computing resource demand feature vector, the current computing resource allocation feature vector, and the residual computing resource feature vector to obtain a corrected computing resource demand feature vector, a corrected current computing resource allocation feature vector, and a corrected residual computing resource feature vector, respectively; and a bayesian computing unit 292, configured to calculate a position-by-position point product of the corrected current computing resource allocation feature vector and the corrected remaining computing resource feature vector, and divide the position-by-position point product by the computing resource demand feature vector to obtain the classification feature vector, where the corrected current computing resource allocation feature vector is used as a prior probability vector, the corrected remaining computing resource feature vector is used as an event probability vector, and the computing resource demand feature vector is used as an evidence probability vector.
Specifically, in this embodiment of the present application, the allocation plan generating module 300 is configured to pass the classification feature vector through a classifier to obtain a classification result, where the classification result is used to represent the count of the server at the current time pointCalculating whether the resource allocation proportion needs to be adjusted. In a specific example, the allocation plan generating module is further configured to: processing the classification feature vector using the classifier to obtain the classification result with the following formula:
Figure DEST_PATH_IMAGE009
wherein, W 1 To W n As a weight matrix, B 1 To B n Is a bias vector and X is the classification feature vector.
In summary, the intelligent computing resource allocation system 200 for a two-way rack server according to the embodiment of the present application is illustrated, the method carries out implicit correlation characteristic mining on the required computing resource quantity of the client, the residual computing resource quantity of the server and the computing resource quantity distributed by the client in a time dimension through a deep neural network model, the rationality judgment of the allocation proportion of the computing resources of the server is carried out by integrating the feature information of the three, in the fusion process of the characteristics, the excessive fragmentation of the decision area in the classification target domain caused by the characteristic value set due to the characteristic value outside the distribution of the set is avoided by carrying out the structural understanding of the information rule on the characteristic value and the class condition to which the characteristic value belongs, and then, the constraint that each feature is distributed to a probabilistic classification target is realized, so that the classification accuracy is improved.
As described above, the intelligent computing resource allocation system 200 of a two-way rack server according to the embodiment of the present application can be implemented in various terminal devices, such as a server of an intelligent computing resource allocation algorithm of a two-way rack server. In one example, the intelligent computing resource allocation system 200 of a two-way rack server according to an embodiment of the present application may be integrated into a terminal device as one software module and/or hardware module. For example, the intelligent computing resource allocation system 200 of the two-way rack server may be a software module in the operating system of the terminal device, or may be an application developed for the terminal device; of course, the intelligent computing resource distribution system 200 of the two-way rack server can also be one of many hardware modules of the terminal device.
Alternatively, in another example, the intelligent computing resource distribution system 200 of the two-way rack server and the terminal device may be separate devices, and the intelligent computing resource distribution system 200 of the two-way rack server may be connected to the terminal device through a wired and/or wireless network and transmit the interaction information according to an agreed data format.
Exemplary method
FIG. 3 illustrates a flow chart of a method for allocating computing resources of the intelligent allocation system of two-way rack servers. As shown in fig. 3, the allocation method of the computing resource intelligent allocation system of a two-way rack server according to the embodiment of the present application includes the steps of: s110, acquiring computing resource requests of a plurality of clients at a plurality of preset time points including a current time point, wherein the computing resource requests comprise required computing resource amount; s120, inputting the required computing resource amount in the computing resource requests of the plurality of clients at each preset time point into a context encoder to obtain a global demand feature vector corresponding to each preset time point; s130, calculating a global mean value of the global demand characteristic vector corresponding to each preset time point to obtain a global calculation demand characteristic value corresponding to each preset time point; s140, arranging the global computation demand characteristic values corresponding to the preset time points into input vectors according to the time dimension, and then obtaining computation resource demand characteristic vectors through a first time sequence encoder comprising a one-dimensional convolutional layer; s150, acquiring the residual computing resource amount of the server at a plurality of preset time points including the current time point; s160, the residual computing resource amount of the server at the plurality of preset time points including the current time point is processed by a second time sequence encoder including a one-dimensional convolution layer to obtain a residual computing resource feature vector; s170, acquiring the amount of computing resources distributed to the plurality of clients by the server at the current time point; s180, enabling the computing resource amount distributed to the plurality of clients by the server at the current time point to pass through a third time sequence encoder comprising a one-dimensional convolutional layer to obtain a current computing resource distribution characteristic vector; s190, fusing the computing resource demand characteristic vector, the current computing resource allocation characteristic vector and the residual computing resource characteristic vector by using a Bayesian probability model to obtain a classification characteristic vector; and S200, the classified feature vectors are processed by a classifier to obtain a classification result, and the classification result is used for indicating whether the calculation resource distribution proportion of the server at the current time point needs to be adjusted or not.
FIG. 4 is a schematic diagram illustrating an architecture of an allocation method of a computing resource intelligent allocation system of a two-way rack server according to an embodiment of the present application. As shown in fig. 4, in the network architecture of the allocation method of the intelligent computing resource allocation system for two-way rack-mounted servers, first, the obtained required computing resource amount (e.g., P1 shown in fig. 4) in the computing resource requests of the clients at each of the predetermined time points is input to a context encoder (e.g., E shown in fig. 4) to obtain a global demand feature vector (e.g., VF shown in fig. 4) corresponding to each of the predetermined time points; then, calculating a global mean value of the global demand feature vectors corresponding to the predetermined time points to obtain global calculated demand feature values (for example, CV as illustrated in fig. 4) corresponding to the predetermined time points; then, arranging the global computation demand characteristic values corresponding to the preset time points into an input vector (for example, V as illustrated in FIG. 4) according to a time dimension, and then passing the input vector through a first time sequence encoder (for example, E1 as illustrated in FIG. 4) comprising one-dimensional convolutional layers to obtain a computation resource demand characteristic vector (for example, VF1 as illustrated in FIG. 4); then, passing the obtained remaining computing resource amount (e.g., P2 as illustrated in fig. 4) of the server at a plurality of predetermined time points including the current time point through a second time-series encoder (e.g., E2 as illustrated in fig. 4) including one-dimensional convolutional layers to obtain a remaining computing resource feature vector (e.g., VF2 as illustrated in fig. 4); then, passing the obtained amount of computing resources (e.g., P3 as illustrated in fig. 4) allocated by the server to the plurality of clients at the current time point through a third time-sequence encoder (e.g., E3 as illustrated in fig. 4) containing one-dimensional convolutional layers to obtain a current computing resource allocation eigenvector (e.g., VF3 as illustrated in fig. 4); then, fusing the computing resource demand feature vector, the current computing resource allocation feature vector and the remaining computing resource feature vector using a bayesian probabilistic model to obtain a classification feature vector (e.g., VC as illustrated in fig. 4); and finally, passing the classification feature vector through a classifier (e.g., a circle S as illustrated in fig. 4) to obtain a classification result, wherein the classification result is used for indicating whether the allocation proportion of the computing resources of the server at the current time point needs to be adjusted.
More specifically, in steps S110 and S120, computing resource requests of a plurality of clients at a plurality of predetermined time points including a current time point are obtained, the computing resource requests include a required computing resource amount, and the required computing resource amount in the computing resource requests of the plurality of clients at each of the predetermined time points is input to a context encoder to obtain a global demand feature vector corresponding to each of the predetermined time points. It should be understood that, in the existing solution allocation of computing resources of a server, it usually allocates computing resources based on the computing requirements of clients, a mode of allocating resources preferentially by first-come users is usually adopted, but this allocation mode may result in that other clients cannot obtain enough computing resources when sharing the server, and the server itself also needs to maintain normal operation, such as data transmission, data storage, data coordination, that is, when the server allocates computing resources, in addition to considering the requirements of all the clients for computing resources as a whole, the remaining computing resources of the server itself also need to be considered. Therefore, in the technical scheme of the application, the mining of the association characteristics is performed by using the deep neural network model from the amount of the required computing resources of the client, the amount of the remaining computing resources of the server and the amount of the computing resources allocated by the client, and then the classification is performed by using the classifier so as to obtain the rationality of the computing resource allocation proportion of the server. Thus, the loss of resources can be avoided under the condition of ensuring that the computing resources of the server are reasonably distributed.
That is, specifically, in the technical solution of the present application, first, computing resource requests of a plurality of clients at a plurality of predetermined time points including a current time point are obtained, where the computing resource requests include a required amount of computing resources. It should be understood that, in consideration of the correlation between the required computing resource amounts in the computing resource requests of the clients at the predetermined time points, in the technical solution of the present application, a context encoder, for example, based on a Bert model of a converter, is used to encode the required computing resource amounts in the computing resource requests of the clients at the predetermined time points, so as to extract a global-based high-dimensional semantic feature between the required computing resource amounts to better characterize the computing resource requirement features of the clients, thereby obtaining a plurality of client feature vectors. In this way, the plurality of client-side feature vectors are cascaded to obtain the global demand feature vector corresponding to each predetermined time point in the time dimension.
More specifically, in steps S130 and S140, a global mean of the global demand feature vectors corresponding to the predetermined time points is calculated to obtain global calculation demand feature values corresponding to the predetermined time points, and the global calculation demand feature values corresponding to the predetermined time points are arranged as an input vector according to a time dimension, and then pass through a first time-sequence encoder including a one-dimensional convolutional layer to obtain a calculation resource demand feature vector. That is, in the technical solution of the present application, further, a global mean of the global demand feature vector corresponding to each of the predetermined time points may be calculated to obtain a global calculation demand feature value having global sample dimension associated feature information corresponding to each of the predetermined time points. Then, considering that the global computation requirement characteristic value of each predetermined time point has a dynamic change characteristic in the time dimension, in order to extract such a change characteristic in the time dimension, the global computation requirement characteristic values corresponding to each predetermined time point are further arranged as an input vector according to the time dimension, so as to obtain a computation resource requirement characteristic vector by a first time sequence encoder including a one-dimensional convolutional layer after integrating the characteristic information of each sample in the time dimension. Accordingly, in a specific example, the time-series encoder is composed of full-connected layers and one-dimensional convolutional layers which are alternately arranged, and extracts the association of each global computation requirement characteristic value in a time-series dimension through one-dimensional convolutional coding and extracts the high-dimensional implicit characteristic of each global computation requirement characteristic value through full-connected coding.
More specifically, in step S150 and step S160, the remaining computing resource amounts of the servers at the predetermined time points including the current time point are obtained, and the remaining computing resource amounts of the servers at the predetermined time points including the current time point are passed through a second time-series encoder including one-dimensional convolutional layers to obtain a remaining computing resource feature vector. It should be understood that, in the technical solution of the present application, in order to determine the rationality of the allocation of the computing resources based on the three of the required computing resource amount of the client, the remaining computing resource amount of the server, and the computing resource amount allocated by the client, it is necessary to perform deep feature mining on the associated feature information of the three, and for the remaining computing resource amount of the server, first, the remaining computing resource amounts of the server at a plurality of predetermined time points including the current time point are obtained. Then, considering the remaining computing resource amount of the server in the time sequence dimension, the server has associated implicit feature information with dynamics in time, so in the technical scheme of the application, a time sequence encoder is adopted to extract the implicit dynamic associated features of the server in the time sequence, and accordingly, the remaining computing resource feature vector is obtained.
More specifically, in step S170 and step S180, the amount of computing resources allocated by the server to the plurality of clients at the current time point is obtained, and the amount of computing resources allocated by the server to the plurality of clients at the current time point is passed through a third sequential encoder including a one-dimensional convolutional layer to obtain a current computing resource allocation eigenvector. It should be understood that, in the technical solution of the present application, in order to determine the rationality of the allocation of the computing resources based on the three of the required computing resource amount of the client, the remaining computing resource amount of the server, and the computing resource amount allocated by the client, it is necessary to perform deep feature mining on the associated feature information of the three, and for the computing resource amount allocated by the client, first, the computing resource amount allocated by the server to the plurality of clients at the current time point is obtained. Then, considering that the amount of computing resources allocated to the plurality of clients by the server at the current time point in the time sequence dimension has associated implicit feature information that is dynamic in time, in the technical scheme of the present application, a time sequence encoder is adopted to extract the implicit dynamic associated features of the server at the time sequence, so as to obtain the current computing resource allocation feature vector.
More specifically, in step S190, a bayesian probabilistic model is used to fuse the computing resource demand feature vector, the current computing resource allocation feature vector and the remaining computing resource feature vector to obtain a classification feature vector. It should be appreciated that further, in view of using the current computing resource allocation feature vector as a prior probability, the technical solution in the present application aims to update the prior probability to obtain a posterior probability when there is new evidence, i.e. when there are new computing resource requests of the plurality of clients. Then, according to a bayesian formula, a posterior probability is a prior probability multiplied by an event probability divided by an evidence probability, and therefore, in the technical scheme of the present application, a bayesian probability model is used to fuse the computing resource demand feature vector, the current computing resource allocation feature vector and the residual computing resource feature vector to obtain a classification feature vector, wherein the current computing resource allocation feature vector is used as a prior, the residual computing resource feature vector is used as an event, and the computing resource demand feature vector is used as an evidence. In this way, the classification feature vector can be passed through a classifier to obtain a classification result indicating whether the allocation proportion of the computing resources of the server at the current time point needs to be adjusted. It should be understood that by performing the structural understanding of the information rules on the eigenvalues and the class conditions to which the eigenvalues belong, excessive fragmentation of the decision region of the eigenvalue set in the classification target domain due to the out-of-distribution eigenvalues of the set is avoided, so that the eigenvalue set of the eigenvector expresses the eigenvalue distribution with a robust conditioned class boundary, thereby realizing the constraint of each eigenvalue distribution to the probabilistic classification target, improving the classification effect of the classification eigenvector obtained by the bayesian probability model, and further improving the accuracy of the rationality judgment of the calculation resource distribution proportion of the server at the current time point.
More specifically, in step S200, the classified feature vector is passed through a classifier to obtain a classification result, and the classification result is used to indicate whether the allocation ratio of the computing resources of the server at the current time point needs to be adjusted.
In summary, the allocation method of the computing resource intelligent allocation system of the two-way rack server according to the embodiment of the present application is illustrated, the method carries out implicit correlation characteristic mining on the required computing resource quantity of the client, the residual computing resource quantity of the server and the computing resource quantity distributed by the client in a time dimension through a deep neural network model, the rationality judgment of the allocation proportion of the computing resources of the server is carried out by integrating the feature information of the three, in the fusion process of the characteristics, the excessive fragmentation of the decision area in the classification target domain caused by the characteristic value set due to the characteristic value outside the distribution of the set is avoided by carrying out the structural understanding of the information rule on the characteristic value and the class condition to which the characteristic value belongs, and then, the constraint that each feature is distributed to a probabilistic classification target is realized, so that the classification accuracy is improved.

Claims (7)

1. A computing resource intelligent distribution system of a two-way rack-mounted server is characterized by comprising:
the client computing request acquisition module is used for acquiring computing resource requests of a plurality of clients at a plurality of preset time points including the current time point, wherein the computing resource requests comprise the required computing resource amount;
the computing request data encoding module is used for inputting the required computing resource quantity in the computing resource requests of the clients at each preset time point into a context encoder so as to obtain a global demand characteristic vector corresponding to each preset time point;
the global mean processing module is used for calculating the global mean of the global demand characteristic vector corresponding to each preset time point to obtain a global calculation demand characteristic value corresponding to each preset time point;
the calculation request time sequence coding module is used for arranging the global calculation demand characteristic values corresponding to the preset time points into input vectors according to the time dimension and then obtaining calculation resource demand characteristic vectors through a first time sequence coder comprising a one-dimensional convolutional layer;
the residual computing resource quantity acquisition module is used for acquiring the residual computing resource quantity of the server at a plurality of preset time points including the current time point;
the residual computing resource amount coding module is used for enabling the residual computing resource amount of the server at a plurality of preset time points including the current time point to pass through a second time sequence encoder comprising a one-dimensional convolutional layer so as to obtain a residual computing resource feature vector;
the distribution data acquisition module is used for acquiring the amount of computing resources distributed to the plurality of clients by the server at the current time point;
the distribution data encoding module is used for enabling the computing resource amount distributed to the plurality of clients by the server at the current time point to pass through a third time sequence encoder containing a one-dimensional convolution layer so as to obtain a current computing resource distribution characteristic vector;
a Bayesian fusion module for fusing the computing resource demand feature vector, the current computing resource allocation feature vector and the residual computing resource feature vector using a Bayesian probability model to obtain a classification feature vector; and
the distribution scheme generation module is used for enabling the classification characteristic vectors to pass through a classifier to obtain a classification result, and the classification result is used for indicating whether the calculation resource distribution proportion of the server at the current time point needs to be adjusted or not;
wherein, the Bayesian fusion module comprises:
a feature distribution correction unit, configured to perform class-based condition boundary constraint on the computing resource demand feature vector, the current computing resource allocation feature vector, and the remaining computing resource feature vector to obtain a corrected computing resource demand feature vector, a corrected current computing resource allocation feature vector, and a corrected remaining computing resource feature vector; and
a Bayesian calculating unit, configured to calculate a position-point-by-position product of the corrected current computing resource allocation feature vector and the corrected residual computing resource feature vector, and divide the position-by-position product by the computing resource demand feature vector to obtain the classification feature vector, where the corrected current computing resource allocation feature vector is used as a prior probability vector, the corrected residual computing resource feature vector is used as an event probability vector, and the computing resource demand feature vector is used as an evidence probability vector;
wherein the feature distribution correction unit is configured to: performing class-condition-based boundary constraint on the computing resource demand characteristic vector, the current computing resource allocation characteristic vector and the residual computing resource characteristic vector respectively according to the following formulas to obtain the corrected computing resource demand characteristic vector, the corrected current computing resource allocation characteristic vector and the corrected residual computing resource characteristic vector;
wherein the formula is:
Figure 37514DEST_PATH_IMAGE001
whereinf i1 、f i2 Andf i3 the computing resource demand eigenvector, the current computing resource allocation eigenvector, and the remaining computing resource eigenvector, respectivelyA characteristic value of each position of the quantity, andf i1 '、f i2 ' andf i3 ' are the eigenvalues of each position of the corrected computing resource demand eigenvector, the corrected current computing resource allocation eigenvector and the corrected remaining computing resource eigenvector, respectively;
wherein the Bayesian computation unit is configured to: taking the corrected current computing resource allocation feature vector as a prior probability vector, the corrected residual computing resource feature vector as an event probability vector and the computing resource demand feature vector as an evidence probability vector, and calculating the position-point-by-position product of the corrected current computing resource allocation feature vector and the corrected residual computing resource feature vector by the following formula and then dividing the product by the computing resource demand feature vector to obtain the classification feature vector;
wherein the formula is:
Figure 766436DEST_PATH_IMAGE002
whereinf i2 ' is the value of each position in the corrected current computing resource allocation feature vector,f i3 ' andf i1 ' is a value for each position in the corrected remaining computing resource feature vector and the corrected computing resource demand feature vector, respectively, andf is ' is the value of each location in the classification feature vector.
2. The intelligent computing resource allocation system for a two-way rack server of claim 1, wherein the computing request data encoding module comprises:
a context encoding unit, configured to perform global context-based semantic encoding on the required amount of computing resources in the computing resource requests of the multiple clients at each of the predetermined time points using a Bert model based on a converter of the context encoder to obtain multiple client feature vectors; and
and the cascading unit is used for cascading the plurality of client characteristic vectors to obtain the global demand characteristic vector corresponding to each preset time point.
3. The intelligent computing resource allocation system for a two-way rack server of claim 2, wherein the computing request timing encoding module comprises:
the first arrangement unit is used for arranging the global calculation demand characteristic values corresponding to the preset time points into input vectors according to time dimensions;
a first full-concatenation coding unit, configured to perform full-concatenation coding on the input vector using a full-concatenation layer of the first timing coder according to a following formula to extract a high-dimensional implicit feature of a feature value of each position in the input vector, where the formula is:
Figure 376409DEST_PATH_IMAGE003
where X is the input vector, Y is the output vector, W is the weight matrix, B is the offset vector,
Figure 874386DEST_PATH_IMAGE004
represents a matrix multiplication;
a first one-dimensional convolution coding unit, configured to perform one-dimensional convolution coding on the input vector by using a one-dimensional convolution layer of the first time sequence encoder according to the following formula to extract a high-dimensional implicit correlation feature between feature values of each position in the input vector, where the formula is:
Figure 646033DEST_PATH_IMAGE005
wherein the content of the first and second substances,ais a convolution kernelxWidth in the direction,FIs a convolution kernel parameter vector,GIs a matrix of local vectors operating with a convolution kernel,wthe size of the convolution kernel.
4. The intelligent computing resource allocation system for a two-way rack-mounted server of claim 3, wherein the remaining computing resource amount encoding module comprises:
the second arrangement unit is used for arranging the residual computing resource quantity of the servers at a plurality of preset time points including the current time point into residual input vectors according to the time dimension;
a second full-concatenation coding unit, configured to perform full-concatenation coding on the residual input vector using a full-concatenation layer of the second time sequence encoder according to a following formula to extract a high-dimensional implicit feature of a feature value at each position in the residual input vector, where the formula is:
Figure 127830DEST_PATH_IMAGE003
where X is the input vector, Y is the output vector, W is the weight matrix, B is the offset vector,
Figure 10335DEST_PATH_IMAGE004
represents a matrix multiplication;
a second one-dimensional convolution coding unit, configured to perform one-dimensional convolution coding on the residual input vector by using a one-dimensional convolution layer of the second time-series encoder according to the following formula to extract a high-dimensional implicit correlation feature between feature values of each position in the residual input vector, where the formula is:
Figure 362819DEST_PATH_IMAGE006
wherein the content of the first and second substances,ais a convolution kernelxA width in the direction,FIs a convolution kernel parameter vector,GIs a matrix of local vectors operating with a convolution kernel,wthe size of the convolution kernel.
5. The intelligent computing resource allocation system for a two-way rack server of claim 4, wherein the allocation data encoding module comprises:
a third arrangement unit, configured to arrange, according to a time dimension, the amount of computing resources allocated by the server to the plurality of clients as an allocation input vector;
a third full-concatenation coding unit, configured to perform full-concatenation coding on the allocation input vector by using a full-concatenation layer of the third time sequence encoder according to a following formula to extract a high-dimensional implicit feature of a feature value at each position in the allocation input vector, where the formula is:
Figure 836526DEST_PATH_IMAGE003
where X is the input vector, Y is the output vector, W is the weight matrix, B is the offset vector,
Figure 274461DEST_PATH_IMAGE004
represents a matrix multiplication;
a third one-dimensional convolutional coding unit, configured to perform one-dimensional convolutional coding on the allocation input vector using a one-dimensional convolutional layer of the third time-series encoder according to the following formula to extract high-dimensional implicit correlation features between feature values of each position in the allocation input vector, where the formula is:
Figure 475504DEST_PATH_IMAGE007
wherein the content of the first and second substances,ais a convolution kernel inxA width in the direction,FIs a convolution kernel parameter vector,GIs a matrix of local vectors operating with a convolution kernel,wthe size of the convolution kernel.
6. The intelligent computing resource allocation system for a two-way rack server of claim 5, wherein the allocation plan generation module is further configured to: processing the classification feature vector using the classifier to obtain the classification result with the following formula:
Figure 948073DEST_PATH_IMAGE008
wherein W is 1 To W n As a weight matrix, B 1 To B n Is a bias vector and X is the classification feature vector.
7. A method for distributing computing resource intelligent distribution system of a two-way rack-mounted server is characterized by comprising the following steps:
acquiring computing resource requests of a plurality of clients at a plurality of preset time points including a current time point, wherein the computing resource requests comprise required computing resource amount;
inputting the required computing resource amount in the computing resource requests of the plurality of clients at each preset time point into a context encoder to obtain a global demand characteristic vector corresponding to each preset time point;
calculating a global mean value of the global demand characteristic vector corresponding to each preset time point to obtain a global calculation demand characteristic value corresponding to each preset time point;
arranging the global computation demand characteristic values corresponding to the preset time points into input vectors according to the time dimension, and then obtaining computation resource demand characteristic vectors through a first time sequence encoder comprising a one-dimensional convolutional layer;
acquiring the residual computing resource amount of the server at a plurality of preset time points including the current time point;
the residual computing resource amount of the server at a plurality of preset time points including the current time point is processed by a second time sequence encoder comprising a one-dimensional convolution layer to obtain a residual computing resource feature vector;
acquiring the amount of computing resources distributed to the plurality of clients by the server at the current time point;
the computing resource amount distributed to the plurality of clients by the server at the current time point is processed by a third time sequence encoder containing a one-dimensional convolutional layer to obtain a current computing resource distribution characteristic vector;
fusing the computing resource demand feature vector, the current computing resource allocation feature vector and the residual computing resource feature vector by using a Bayesian probability model to obtain a classification feature vector; and
the classified feature vectors are processed by a classifier to obtain a classification result, and the classification result is used for indicating whether the calculation resource allocation proportion of the server at the current time point needs to be adjusted or not;
wherein the fusing the computing resource demand feature vector, the current computing resource allocation feature vector, and the remaining computing resource feature vector using a Bayesian probability model to obtain a classification feature vector comprises:
respectively carrying out class-condition-based boundary constraint on the computing resource demand characteristic vector, the current computing resource allocation characteristic vector and the residual computing resource characteristic vector to obtain a corrected computing resource demand characteristic vector, a corrected current computing resource allocation characteristic vector and a corrected residual computing resource characteristic vector; and
taking the corrected current computing resource distribution feature vector as a prior probability vector, the corrected residual computing resource feature vector as an event probability vector and the computing resource demand feature vector as an evidence probability vector, and calculating the position-point-by-position product of the corrected current computing resource distribution feature vector and the corrected residual computing resource feature vector and dividing the product by the computing resource demand feature vector to obtain the classification feature vector;
wherein the respectively performing class-based conditional boundary constraint on the computing resource demand feature vector, the current computing resource allocation feature vector, and the residual computing resource feature vector to obtain a corrected computing resource demand feature vector, a corrected current computing resource allocation feature vector, and a corrected residual computing resource feature vector comprises: respectively carrying out class condition boundary-based constraint on the computing resource demand characteristic vector, the current computing resource allocation characteristic vector and the residual computing resource characteristic vector according to the following formulas to obtain the corrected computing resource demand characteristic vector, the corrected current computing resource allocation characteristic vector and the corrected residual computing resource characteristic vector;
wherein the formula is:
Figure 61523DEST_PATH_IMAGE009
whereinf i1 、f i2 Andf i3 is a feature value of each position of the computing resource demand feature vector, the current computing resource allocation feature vector and the remaining computing resource feature vector, respectively, andf i1 '、f i2 ' andf i3 ' are the eigenvalues of each position of the corrected computing resource demand eigenvector, the corrected current computing resource allocation eigenvector and the corrected remaining computing resource eigenvector, respectively;
wherein, the calculating the location-point-by-location multiplication of the corrected current computing resource allocation feature vector and the corrected residual computing resource feature vector by the computing resource demand feature vector to obtain the classification feature vector by using the corrected current computing resource allocation feature vector as a prior probability vector, the corrected residual computing resource feature vector as an event probability vector, and the computing resource demand feature vector as an evidence probability vector comprises: taking the corrected current computing resource allocation feature vector as a prior probability vector, the corrected residual computing resource feature vector as an event probability vector and the computing resource demand feature vector as an evidence probability vector, and calculating the position-point-by-position product of the corrected current computing resource allocation feature vector and the corrected residual computing resource feature vector by the following formula and then dividing the product by the computing resource demand feature vector to obtain the classification feature vector;
wherein the formula is:
Figure 517912DEST_PATH_IMAGE010
whereinf i2 ' is the value of each position in the corrected current computing resource allocation feature vector,f i3 ' andf i1 ' is a value for each position in the corrected remaining computing resource feature vector and the corrected computing resource demand feature vector, respectively, andf is ' is the value of each location in the classification feature vector.
CN202210810557.6A 2022-07-11 2022-07-11 Intelligent computing resource distribution system and method for double-path rack-mounted server Active CN114860462B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210810557.6A CN114860462B (en) 2022-07-11 2022-07-11 Intelligent computing resource distribution system and method for double-path rack-mounted server

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210810557.6A CN114860462B (en) 2022-07-11 2022-07-11 Intelligent computing resource distribution system and method for double-path rack-mounted server

Publications (2)

Publication Number Publication Date
CN114860462A CN114860462A (en) 2022-08-05
CN114860462B true CN114860462B (en) 2022-09-30

Family

ID=82626370

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210810557.6A Active CN114860462B (en) 2022-07-11 2022-07-11 Intelligent computing resource distribution system and method for double-path rack-mounted server

Country Status (1)

Country Link
CN (1) CN114860462B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116760772B (en) * 2023-08-21 2023-10-27 北京中科网芯科技有限公司 Control system and method for converging flow divider
CN117116498A (en) * 2023-10-23 2023-11-24 吉林大学 Mobile ward-round data processing system and method thereof

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105068874B (en) * 2015-08-12 2018-11-30 国家电网公司 A kind of on-demand dynamic allocation method of resource of combination Docker technology
CN110858161B (en) * 2018-08-24 2023-05-12 阿里巴巴集团控股有限公司 Resource allocation method, device, system, equipment and medium
CN109376006B (en) * 2018-09-04 2021-09-21 西安电子科技大学 Resource integration method based on time-varying characteristics of user requirements in cloud computing environment
CN110032447B (en) * 2019-04-11 2022-02-22 北京百度网讯科技有限公司 Method and apparatus for allocating resources
CN110138612B (en) * 2019-05-15 2020-09-01 福州大学 Cloud software service resource allocation method based on QoS model self-correction
CN114677185B (en) * 2022-05-26 2022-10-18 深圳市虎瑞科技有限公司 Intelligent large-screen advertisement intelligent recommendation system and recommendation method thereof

Also Published As

Publication number Publication date
CN114860462A (en) 2022-08-05

Similar Documents

Publication Publication Date Title
CN114860462B (en) Intelligent computing resource distribution system and method for double-path rack-mounted server
CN112860695B (en) Monitoring data query method, device, equipment, storage medium and program product
KR102555607B1 (en) Method and apparatus for annotating data, device, storage medium and computer program
CN112214316A (en) Data resource allocation method based on Internet of things and cloud computing server
CN110096697B (en) Word vector matrix compression method and device, and method and device for obtaining word vectors
CN115879542A (en) Federal learning method oriented to non-independent same-distribution heterogeneous data
Gao et al. A deep learning framework with spatial-temporal attention mechanism for cellular traffic prediction
CN112508408A (en) Mapping model construction method of wireless resource management index under edge calculation
CN112488563A (en) Determination method and device for force calculation parameters
CN116302481A (en) Resource allocation method and system based on sparse knowledge graph link prediction
CN114329127B (en) Feature binning method, device and storage medium
CN113052629B (en) Network user image drawing method based on CECU system intelligent algorithm model
CN113064554B (en) Optimal storage node matching method, device and medium based on distributed storage
CN111092755B (en) Edge service migration simulation method based on resource occupation
CN115016889A (en) Virtual machine optimization scheduling method for cloud computing
CN103645889B (en) Dynamic software self-adaption generating method
CN113438678A (en) Method and device for distributing cloud resources for network slices
CN113760484A (en) Data processing method and device
CN116561696B (en) Multi-dimensional user adjustable load rapid aggregation method and system thereof
CN111429045B (en) Energy internet clustering method based on region symmetry
CN114079997B (en) High-performance communication method based on WSN (Wireless sensor network) improved routing protocol
CN116540990B (en) Code integration method and device for realizing electronic product based on embedded mode
CN116662415B (en) Intelligent matching method and system based on data mining
CN113938944A (en) Method for evaluating 5G communication technology adaptability of power distribution network
CN115545107A (en) Cloud computing method and system based on mass power data

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant